← Back
Deep Learning labs
XOR, optimisers, CNNs, regularisation, transfer learning and GANs
Problem
Guided course exercises.
Data
MNIST, Fashion-MNIST, tf_flowers.
Methods
- XOR neural network from scratch (NumPy)
- SGD vs Adam vs RMSprop on Fashion-MNIST (PyTorch)
- CNN on MNIST with filters and feature maps
- Normalisation, L2 and Dropout
- VGG16 transfer learning + fine-tuning (tf_flowers)
- GAN on MNIST
Results
- Optimisers (Fashion-MNIST, 10 epochs): Adam and RMSprop ~88% vs 77.23% for SGD.
- Regularisation (Fashion-MNIST, test): baseline 84.88%, normalised 87.57%, L2 + Dropout 86.32% with a smaller generalisation gap.
Tools
PythonPyTorchTensorFlow / KerasMLflowDatabricks